I study how minds, biological and synthetic, coordinate with each other and with the world, and where that coordination . My fields are human augmentation, HCI, cognitive computational neuroscience, computational neurolinguistics, neurorobotics, complex systems, multi-agent systems, mechanistic interpretability, brain computer interfaces, mixed reality, human-machine teaming, collective intelligence, and nanotechnology. What I am building now is on the current work page; how I keep commits reproducible is on engineering. This page is the research underneath: each thread, why I keep at it, and where the evidence lives.
The question from cosmos carries over here: what happens when one mind meets another, biological or synthetic, and neither gets to pretend the other was not in the room.
Neural assemblies
If word learning and working memory can run on sparse, brain-scale circuits instead of a backpropagated transformer, the story about what minds are doing changes. Neural assemblies is my runnable test of it: the Assembly Calculus (Papadimitriou et al., PNAS 2020), a formal model of how populations of neurons bind, store, and recall patterns, plus the NEMO language-acquisition model (Mitropolsky & Papadimitriou, 2025), built with the MIT Poggio Lab (CBMM) and Daniel Mitropolsky. Install it from PyPI as neural-assemblies.
Six operations do most of the work: project, associate, merge, separate, sequence memorize, ordered recall. Each runs in a k-WTA area, where only the k most active cells and the rest stay quiet. That competition is what keeps the representations sparse. The primitives experiments page has the recovery curves, convergence criteria, and full tables. The result I care about most is the N400 formalization. In EEG, the N400 is the bump that appears when a word is hard to fit into context; here its amplitude tracks pre-k-WTA energy(the input to an area before the winners are chosen), not the post-k-WTA readouts everyone else measures. The plots and stats are in the claims index. Reproduce them; if a number fails on your machine, open an issue.
Language acquisition
Language is where the calculus has to meet the world: a child does not learn nouns and verbs from a grammar tagger. If part of speech and simple roles like subject and object can emerge from grounded experience in a brain-like model, that challenges both innate grammar modules and “just scale the LLM.” I keep three implementations at different levels of biological fidelity so a single demo cannot fool me.
EmergentParser routes words by grounding(what you see and do with a word): visual experience tends toward nouns, motor experience toward verbs, with no hardcoded POS labels. NemoLanguageLearner is the GPU-scale line, with lexical areas wired to vision and motor. LanguageLearner is the distributional control: no grounding, only statistics from exposure, so categories have to emerge on their own. The assemblies site has runs, corpora, and the comparison details.
I train new words on a fixed curriculum of ten grounded sentences each (naming, describing, action, question, command, with objects, shared attention between speaker and listener, and clear speech acts). Whether that is enough for stable readout is still open; the curriculum is defined on the site, the learning curves are not finished.
GPU infrastructure
The assembly calculus only matters if you can iterate on it. Each step finds the top-k most active neurons across pools of thousands to millions of cells. Top-k is slow on a GPU, and atomic collisions on shared memory make it slower. I wrote CUDA kernels (warp reduction, radix top-k, memory coalescing) that turned “wait on the job queue” into “test the hypothesis now.” The wall I keep hitting is memory: dense weight matrices grow with the square of area size, so very large n still hurts even when each step is fast.
Benchmarks, engine selection (numpy_sparse, cupy, torch_sparse), and throughput tables live in the assemblies repo and docs. Read there if you need steps per second at a given neuron count n and winner count k.
Mechanistic interpretability
A false-belief task checks whether a system can track what another agent believes, including when that belief is wrong. When a model fails one, is that bad reasoning or a broken internal model of the other mind? Collaboration Circuits is my attempt to answer that inside the weights. The design is pre-registered: hypotheses and analysis written down before the results, so I cannot quietly rearrange the story afterward. It pairs with AI Civilizations, where I study the same coordination failures in groups of models and people.
The experimental plan lists models, stimuli, statistics, and five planned studies. I maintain a fork of circuit-tracer on PyPI for attribution-graph work (which internal features and connections drive a model's answer). The plan in the repo is the running protocol; this section only orients.
TYPHON
Fitting more into one session is long context; remembering across sessions days apart is cross-episode memory. Both need an architecture, not just a bigger window. TYPHON is the harness I am building while the learned memory model is still missing: benchmark, baseline, and a v0 that scores chunks and writes to memory only when a passage looks worth keeping (high surprise, large gradient, or novelty). The target has four layers: local exact recall, fast adaptive memory, sparse episodic traces, slow cross-episode structure. The repo has the current heuristics and eval wiring.
What is still open
About twenty open questions stay on my tracker. These are the ones I return to most:
- Where assembly dynamics hit phase transitions or bifurcations, and the calculus stops being the right story.
- One API across CUDA, CuPy, and NumPy within ten percent of optimal performance (not started).
- at biological parameters: about ten thousand neurons per area, fifty winners, one percent connection probability (N=10,000, K=50, P=0.01), not only the easier ratios already tested.
- Word learning curves with full brain integration in the loop.
- Cross-linguistic generalization beyond English and Russian.
If a claim in the claims index does not reproduce, open an issue on the repo. For surrounding writing, see the papers and the blog. For stance, see philosophy.